Update app.py

app.py

@@ -6,7 +6,7 @@ import json
 import shutil
 from pathlib import Path
 from huggingface_hub import hf_hub_download
-from transformers import pipeline
+from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer
 
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 
@@ -23,49 +23,52 @@ def load_models():
     logging.info("📥 Loading models...")
 
     try:
-        # ====================== English Model ======================
+        # English & Sinhala (unchanged)
         en_repo = "E-motionAssistant/English_LR_Model_New"
         en_vectorizer = joblib.load(hf_hub_download(en_repo, "tfidf_vectorizer.joblib"))
         en_classifier = joblib.load(hf_hub_download(en_repo, "logreg_model.joblib"))
         en_label_encoder = joblib.load(hf_hub_download(en_repo, "label_encoder.joblib"))
 
-        # Load emotion_map.json
         try:
             map_path = hf_hub_download(en_repo, "emotion_map.json")
             with open(map_path, "r", encoding="utf-8") as f:
                 en_emotion_map = json.load(f)
-            logging.info("✅ emotion_map.json loaded for English")
         except:
-            logging.warning("Could not load emotion_map.json")
             en_emotion_map = None
 
-        # ====================== Sinhala Model ======================
         si_vectorizer = joblib.load(hf_hub_download("E-motionAssistant/Sinhala_Text_Emotion_Model_LR", "tfidf_vectorizer.joblib"))
         si_classifier = joblib.load(hf_hub_download("E-motionAssistant/Sinhala_Text_Emotion_Model_LR", "logreg_model.joblib"))
         si_label_encoder = joblib.load(hf_hub_download("E-motionAssistant/Sinhala_Text_Emotion_Model_LR", "label_encoder.joblib"))
 
-        # ====================== TAMIL - FIXED VERSION ======================
-        logging.info("📥 Loading Tamil model...")
+        # ====================== TAMIL - STRONG FIX ======================
+        logging.info("📥 Loading Tamil model with manual components...")
 
-        # Clean old cache (helps fix "always joy" issue)
+        # Clean cache
         try:
             cache_dir = Path.home() / ".cache" / "huggingface" / "hub"
             model_cache = cache_dir / "models--E-motionAssistant--Tamil_Emotion_Recognition_Model"
             if model_cache.exists():
                 shutil.rmtree(model_cache)
-                logging.info("🧹 Cleaned old Tamil model cache")
+                logging.info("🧹 Cleaned Tamil cache")
         except:
             pass
 
+        # Load manually (more reliable than pipeline sometimes)
+        model_name = "E-motionAssistant/Tamil_Emotion_Recognition_Model"
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+        model = AutoModelForSequenceClassification.from_pretrained(model_name)
+
         tamil_pipe = pipeline(
             "text-classification",
-            model="E-motionAssistant/Tamil_Emotion_Recognition_Model",
-            tokenizer="E-motionAssistant/Tamil_Emotion_Recognition_Model",   # Explicit tokenizer
+            model=model,
+            tokenizer=tokenizer,
             device=-1,
             truncation=True,
-            max_length=512
+            max_length=512,
+            top_k=1
         )
-        logging.info("✅ Tamil model + tokenizer loaded successfully")
+
+        logging.info("✅ Tamil model loaded with manual tokenizer & model")
 
         models = (en_vectorizer, en_classifier, en_label_encoder,
                   si_vectorizer, si_classifier, si_label_encoder, tamil_pipe)
@@ -90,7 +93,7 @@ class PredictRequest(BaseModel):
 
 @app.get("/")
 def root():
-    return {"status": "ok", "message": "Emotion Detector API is running"}
+    return {"status": "ok"}
 
 
 @app.post("/predict")
@@ -98,9 +101,7 @@ def predict(req: PredictRequest):
     if not req.text or not req.text.strip():
         return {"error": "Text cannot be empty"}
 
-    # Safety check
     if models is None:
-        logging.warning("Models not loaded. Loading now...")
         load_models()
 
     en_vec, en_clf, en_le, si_vec, si_clf, si_le, tamil_pipe = models
@@ -121,22 +122,16 @@ def predict(req: PredictRequest):
             return {"emotion": str(emotion), "language": "Sinhala"}
 
         elif lang == "tamil":
-            logging.info(f"Tamil input: '{req.text[:150]}...'")
+            logging.info(f"Tamil input: {req.text[:200]}...")
 
-            result = tamil_pipe(req.text, truncation=True, max_length=512)
+            result = tamil_pipe(req.text)
 
-            logging.info(f"Tamil raw output: {result}")
+            logging.info(f"Tamil raw result: {result}")
 
-            # Get top prediction
-            if isinstance(result, list) and len(result) > 0:
-                top = result[0] if isinstance(result[0], dict) else result[0][0]
-                emotion = top["label"]
-                score = round(float(top["score"]), 4)
-            else:
-                emotion = "joy"
-                score = 0.0
+            emotion = result[0]['label']
+            score = round(float(result[0]['score']), 4)
 
-            logging.info(f"Tamil Final Prediction → {emotion} (Confidence: {score})")
+            logging.info(f"Tamil Final → {emotion} ({score})")
 
             return {
                 "emotion": emotion,
@@ -144,9 +139,6 @@ def predict(req: PredictRequest):
                 "language": "Tamil"
             }
 
-        else:
-            return {"error": f"Unsupported language: {req.language}"}
-
     except Exception as e:
-        logging.error(f"Prediction error: {e}")
-        return {"error": "Prediction failed. Please try again."}
+        logging.error(f"Error: {e}")
+        return {"error": str(e)}